基于Adaboost与CNN的木材表面缺陷检测

Wood Surface Defect Detection Based on Adaboost and CNN

针对木材智能下料前期需要快速获取木材表面缺陷的位置以及类别信息的要求,基于数字图像处理技术,提出了一个组合算法用于木材表面缺陷的快速识别与定位。使用了Adaboost级联分类器在图像中提取出木材表面缺陷区域的候选框,有效解决了传统分割方法对于多目标难以处理的问题。使用了具有自学习特征能力的CNN(Convolutional Neural Networks)模型对输入的候选框进行分类,克服了传统分类方法中特征难以选择的不足。基于大量样本对模型进行训练,采用200张多缺陷样本进行测试。试验结果表明:检测的召回率为94%,检测的正确率为99%,分类的准确率为97.9%。试验验证了该算法可以满足木材表面缺陷的定位与分类要求。

In order to obtain the location information and category information of wood surface defects in the early stage of intelligent wood cutting, a combined algorithm for fast identification and location of wood surface defects is proposed based on digital image processing technology for fast recognition and location of wood surface defects. The Adaboost cascade classifier algorithm is used to extract the candidate frames of the wood surface defect region in the image in order to solve the problem that the traditional segmentation method is not effective for multi objects processing. A CNN model with self-learning ability is used to classify the input candidate boxes, so as to overcome the difficulty of selecting features in traditional classification methods. 200 multi-target samples are selected for testing,and the results show that the recall rate is 94%, the accuracy of detection is 99%, and the accuracy of classification is 97.9%. The experimental results show that the algorithm can meet the requirements of localization and classification of wood surface defects.